US6832191B1 - Process for implementing a speech recognizer, the related recognizer and process for speech recognition - Google Patents
Process for implementing a speech recognizer, the related recognizer and process for speech recognition Download PDFInfo
- Publication number
- US6832191B1 US6832191B1 US09/649,632 US64963200A US6832191B1 US 6832191 B1 US6832191 B1 US 6832191B1 US 64963200 A US64963200 A US 64963200A US 6832191 B1 US6832191 B1 US 6832191B1
- Authority
- US
- United States
- Prior art keywords
- recognizer
- transitions
- multilingual
- transition
- language
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Lifetime, expires
Links
- 238000000034 method Methods 0.000 title claims description 42
- 230000008569 process Effects 0.000 title claims description 32
- 239000000463 material Substances 0.000 claims abstract description 40
- 238000012549 training Methods 0.000 claims abstract description 34
- 230000007704 transition Effects 0.000 claims description 85
- 230000006870 function Effects 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 7
- 230000001419 dependent effect Effects 0.000 claims description 3
- 238000007670 refining Methods 0.000 claims description 3
- 238000012545 processing Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 239000013067 intermediate product Substances 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000013518 transcription Methods 0.000 description 2
- 230000035897 transcription Effects 0.000 description 2
- 230000001755 vocal effect Effects 0.000 description 2
- 241000220010 Rhode Species 0.000 description 1
- 238000010521 absorption reaction Methods 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000000205 computational method Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000010287 polarization Effects 0.000 description 1
- 230000003252 repetitive effect Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
- G10L15/063—Training
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
- G10L15/00—Speech recognition
- G10L15/005—Language recognition
Definitions
- This invention addresses the techniques for speech recognition and in particular the issue of implementing a speaker independent automatic multilingual speech recognizer also suitable for languages for which speech training material is scant.
- the invention is therefore targeted at being applied preferentially in situations where the use of automatic vocal recognition systems appears advantageous (such as at airport and rail station transit areas, shows and conferences, automatic vocal message controlled systems and the like for instance) and where recognition assurance is desired even when the speaker talks in a not particularly widespread language and/or in a language for which it is neither easy nor advantageous to collect the speech training material normally required for implementing a speech recognizer in a short time.
- This invention has the purpose of supplying a speech recognizer of the type as specified above and according to the invention such purpose is achieved by means of a process having the features called out specifically in the claims that follow.
- the invention also concerns the recognizer for a given language implemented with this process as well as the multilingual recognizer created as an intermediate product of the process itself.
- the invention also extends to the corresponding process for voice recognition.
- FIG. 1 illustrates at the general level the subsequent steps into which the process according to the invention divides
- FIGS. from 2 to 4 illustrate the implementation of some of the steps shown in the blocks of FIG. 1 in greater detail as flow charts.
- the solution according to the invention firstly includes the implementation as an intermediate product of an automatic multilingual speech recognizer in a situation in which an ample amount of speech material is available for all the languages involved. Such automatic multilingual speech recognizer is subsequently exploited to interpolate a sufficiently robust recognizer for a language for which large information databases are unavailable.
- the acoustic phonetic models used are of the transitory/stationary type. These are the models sometimes referred to as APUCD-CTS (Acoustic Phonetic Unit Context Dependent—Class Transitory Stationary) in juxtaposition to context independent models, also called APUCI (Acoustic Phonetic Unit Context Independent).
- implementation of the recognizer according to the invention involves completion of the steps represented in a general form in blocks 1 to 5 of FIG. 1 .
- a multilingual recognizer model is first generated obtained only with trainable languages (block 2 ), after acquisition of the related training material (block 1 ).
- the model for a new language for which the amount of speech training material is scant (block 3 ) is then interpolated starting from such multilingual recognizer model.
- a recognizer model of the new language (block 5 ) is generated based on the result of the above interpolation and the possible online acquisition of the corresponding speech material (block 4 ). Such model is preferably improved based on the speech material acquired online.
- Such generation action essentially corresponds to blocks 1 and 2 of FIG. 1 and the flow chart of FIG. 2 .
- the upper part of this flow chart includes a series of N processes referred to a plurality of N start languages recognized as trainable languages, that is to say as languages (Italian, Spanish and the like for instance) for which a great amount of speech material is available.
- N start languages recognized as trainable languages, that is to say as languages (Italian, Spanish and the like for instance) for which a great amount of speech material is available.
- first step processing is separate for each individual language so the related initial models are not multilingual but specific for each language.
- An initial step (represented by blocks 11 , . . . , 1 N for the various languages) consists of identifying the context independent (CI) units related to the phonemes of the respective languages.
- Hidden Markov models (HMM) are generated for these units by initializing and training (steps 21 , . . . , 2 N—always with reference to the N languages considered above).
- Context independent models are subsequently used as a starting point for the definition, initialization and training of transitory/stationary type units where context dependent transition units represent transition between one phoneme and the next and the stationary part corresponds to the more stable and central component of implementation of such phonemes that continue to be context independent models.
- Initialization of the transition unit in terms of the stochastic structure of Markov models is important to ensure good convergence of the repetitive training process to a optimum point as described in the paper by G. Micca et al. mentioned previously.
- steps 41 , . . . , 4 N of the flow chart of FIG. 2 involves reference to the language corpus 51 , . . . , 5 N (the project for which is represented by blocks 301 , . . . , 30 N) of the respective language. This is also dependent on data from the related database 31 , . . . , 3 N already used for initializing and training models at steps 21 , . . . , 2 N.
- the second part of the speech recognizer creation process is based on merging the transitions obtained previously and specific for each language into classes. This operation is completed by adopting clustering techniques that no longer take account of the language of origin, but are based on the concept of distance between models and are also directed at solving the problem of under-training of the models of some transitions.
- Block 7 of FIG. 2 represents measurement of the distances between transition units referred to the complex of N initial languages.
- the subsequent merging operation is a procedure in two-steps, represented respectively in blocks 8 a and 8 b.
- the first step ( 8 a ) is based on an initial criterion that starts from the transitions closest in terms of distance and merges them into increasingly big classes only when all the transitions to be merged are at less than a given threshold distance from the other components of the same class.
- the second merger criterion (step 8 b ) is directed at overcoming the problem of “rare” transitions not included in any class based on the criterion adopted at step 8 a .
- Rare transitions are therefore added to one of the classes already created or to a new class created for this purpose. This is accomplished by identifying the transition closest to the rare transition and when such transition has already been added to a class and is trainable enough, by forcing the merger even when the distance requirements accounted for at step 8 a have not been respected. If however the transition in question is close to another transition that does not belong to a class, the two transitions are used to create a new class.
- transition or class into which the rare transition is to be merged is not trainable, other transitions closer to the one in question are looked for and so on until a transition or class satisfying the trainability requirement is found.
- a transition or class is defined as trainable if it appears in the training database a sufficient number of times to obtain a robust model.
- Blocks 9 and 10 represent the steps leading to definition of new transition classes on the one hand and to initialization and training of the classes obtained on the other.
- Block 10 output identifies generation of the multilingual recognizer model identified by block 2 in the diagram of FIG. 1 .
- a characteristic feature of the multilingual recognizer described is the fact that although the trend is to initialize the process by attempting as ample as possible a coverage of all transitions between the various phonemes of a language it cannot be excluded that some words of the vocabulary of a given general recognition application generate new transitions not present in the initial models. This event is less probable as more attention is given in defining the language corpus 51 . . . , 5 N of the training database that must guarantee as ample coverage as possible of all the phonetic phenomena susceptible of being represented in a given language. This on particular holds true for the project of the language corpus represented by blocks 301 , . . . , 30 N. In general, the solution of repeating the process in the presence of new transitions does not appear particularly attractive or advantageous especially if the new transitions make up a limited whole versus total start transitions.
- the same recognizer is used for recognizing another language (hereinafter defined as new) for which one is in “conditions of substantial unavailability of speech material”. This means that the speech material available for this language is very scant or in actual fact practically non-existent.
- the assumption is made that a quantity of speech material sufficient is available to train the context-independent acoustic phonetic units (CI—that are no more than some ten units for one language) but inadequate to train hundreds of transition units.
- CI context-independent acoustic phonetic units
- the related database available is represented by block 100 of FIG. 3 .
- Block 100 ′′ of the same figure represents a test stage during which the speech materials from block 100 ′ is used to check speech recognizer 107 better described later on. The purpose is to achieve refinement of the mapping stage represented by block 104 , also subject to be described better later on.
- Database 100 ′ and its test 100 ′′ are susceptible of being developed for instance by collecting a limited set of speech samples from speakers of the “new” not completely trainable language. For instance, in the case of an application for an airport or a similar transit area it can be envisaged that the above set will be collected from persons accessible at such place and who speak the new language.
- the recognizer according to the invention does not use, at least at the design stage, any acoustic data of the new language but exploits the similarity between the phonological system of the new language and that of the various languages making up the multilingual model obtained during the previous stages.
- the attempt is first made to identify with prior knowledge anyhow available for each phoneme of the new language the most similar one belonging to one of the other languages. For instance, the phoneme represented by the same symbol IPA (International Phonetic Alphabet) or SAMPA (Speech Assessment Methods Phonetic Alphabet) is associated; the latter alphabet is a transcription of the IPA symbols into ASCII symbols non extending from number 7 to number 127 . This transcription was applied for the first time in 1989 to the languages of the European Union and was subsequently extended to other languages as well.
- IPA International Phonetic Alphabet
- SAMPA Speech Assessment Methods Phonetic Alphabet
- step 101 represents in both the step during which (after obtaining correspondence of the sound of the new language and those of the multilingual recognizer making up the result of step 2 of FIGS. 1 and 2) the context-independent (CI) units of the new language are initialized with the models of the units they have been associated to. An albeit limited performance recognizer is obtained in this fashion.
- Transition models are multilingual and present a double degree of generality useful for the purpose set: ample coverage and greater robustness.
- these models guarantee coverage of sounds of the new languages with a degree of precision that increases in proportion to the number N of languages making up the system.
- the models under review represent an ample phonetic and acoustic space versus the space obtainable based on monolingual transitions. These models in fact contain information related to several languages so the probabilities that any sound of the new language finds a sufficiently representative model increase.
- the neo-Latin language of Rumanian for instance is closer to a multilingual model consisting of Italian and Spanish than German, which has a rather different phonological structure from the other languages considered. In this perspective, it is even more evident that by increasing the number of languages used for defining the multilingual model the possibilities of finding sounds close to one of the sounds of the new language increase.
- the steps indicated with references 102 and 103 of the diagram of FIG. 3 correspond respectively to training of the units independently of the context of the new language, based on the information obtainable from database 100 and measurement of the distance between the new context-independent models and their multilingual correspondents.
- non-monolingual transitions can be generated between sounds of the same language not present in any word of the vocabulary of the latter. In both cases transitions are obtained which are not included in the original transitions with which the recognizer was built.
- the strategy adopted to identify a transition included among the original one and susceptible of being replaced instead of the missing transition is the same one as described previously for the replacement of new transitions that can be created when the recognition vocabulary is increased.
- a minimum of speech material is required to try out associations made and check the degree of recognition of the words of the languages trained, by changing choices made if polarization of errors towards certain sounds is noted. Once the system is operated, performance can be improved by using the speech material naturally produced by system use by mother tongue speakers.
- This speech material acquisition function is directed at making up a database for the new language, identified with 108 .
- Database 108 is susceptible of being used for a training function (block 109 ) directed at further refining the model of the new language jointly with a multilingual database 110 . All this is to originate an improved model of the new language (identified by reference 5 in both flow charts of FIGS. 3 and 4 and in the diagram of FIG. 1 ).
- the online speech material acquisition function represented by block 4 interacts directly (practically in real time and anyhow during system operation) with database 100 of the new language by acting on database 108 that is continuously updated as a result of the online acquisition of speech material. This possibility is particularly advantageous when the system is used to enable collection of a quantity of speech material sufficient to train models based in the criterion of retroaction described.
- the speech material is enough for training at least context-independent units, only transitions are initialized with models not belonging to the new language. Moreover, measuring the distance between the models of the new language and the multilingual one so that replacement is more precise performs identification of the transitions to be replaced.
- Subsequent training can be targeted at strengthening the recognizer in the new language, which is obtained by weighting the information obtained from the reduced amount of speech materials available for the new language more than the information available for the other N languages.
- This approach is based on the introduction of a parameter ⁇ (with 0 ⁇ 1) that weighs with a value (1 ⁇ ) the information brought from the N languages of the initial multilingual model and with ⁇ the information conveyed by the speech material of the new language.
- This weighting function (block 111 ) is connected to methods already applied for adapting speech recognition system to the speaker and/or the channel.
- Cf. for instance works by Q. Huo, C. H. Lee: “Online adaptive learning of the continuous density hidden Markov model based on approximate recursive Bayes estimate”, IEEE Transactions on Speech and Audio Processing, Volume 5, Number 2, pages 161-172, 1997 and by C. Vair, L. Fissore: “Speaker adaptation of CDHMMs using Bayesian learning” and in “Computational Methods of Speech Pattern Processing”, NATO-ASI Series F: Computer and Systems Sciences, Volume 169, Springer-Verlag, pages 78-83, 1999.
- the model of the new language M ln obtained with this adaptation operation is the result of optimization with respect to parameter ⁇ that can be represented with the following formula
- M ml is the initial multilingual model
- I ln is the information extractable from acoustic data acquired in the new language
- M ln is the resultant model for the new language referable to block 5 of FIGS. 1, 3 and 4 .
- Parameter ⁇ can be optimized in agreement with the formulae or re-estimation of the parameters of HMM models of the MAP (Maximum A Posteriori) technique described in the works mentioned.
- the model of recognizer 5 obtained according to the procedures described previously can be incorporated into any multilingual automatic speech recognition system for the completion of any function susceptible of being competed by such recognizer.
- the related speech recognition process is therefore implemented based on this recognizer, with the possibility (described in particular with reference to block 4 ) of having such process also include online acquisition of the speech material for the new language.
Abstract
Description
Claims (17)
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
ITTO99A0734 | 1999-09-02 | ||
IT1999TO000734A IT1310154B1 (en) | 1999-09-02 | 1999-09-02 | PROCEDURE FOR MAKING A VOICE RECOGNIZER, RELATED RECOGNIZER AND PROCEDURE FOR VOICE RECOGNITION |
Publications (1)
Publication Number | Publication Date |
---|---|
US6832191B1 true US6832191B1 (en) | 2004-12-14 |
Family
ID=11418041
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US09/649,632 Expired - Lifetime US6832191B1 (en) | 1999-09-02 | 2000-08-28 | Process for implementing a speech recognizer, the related recognizer and process for speech recognition |
Country Status (6)
Country | Link |
---|---|
US (1) | US6832191B1 (en) |
EP (1) | EP1081681B1 (en) |
JP (1) | JP2001109490A (en) |
CA (1) | CA2317231C (en) |
DE (1) | DE60021666T2 (en) |
IT (1) | IT1310154B1 (en) |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20020111805A1 (en) * | 2001-02-14 | 2002-08-15 | Silke Goronzy | Methods for generating pronounciation variants and for recognizing speech |
US20030040907A1 (en) * | 2001-08-24 | 2003-02-27 | Sen-Chia Chang | Speech recognition system |
US20050055197A1 (en) * | 2003-08-14 | 2005-03-10 | Sviatoslav Karavansky | Linguographic method of compiling word dictionaries and lexicons for the memories of electronic speech-recognition devices |
US20050171766A1 (en) * | 2002-02-28 | 2005-08-04 | Dario Albesano | Method for accelerating the execution of speech recognition neural networks and the related speech recognition device |
US7406408B1 (en) * | 2004-08-24 | 2008-07-29 | The United States Of America As Represented By The Director, National Security Agency | Method of recognizing phones in speech of any language |
US20130151254A1 (en) * | 2009-09-28 | 2013-06-13 | Broadcom Corporation | Speech recognition using speech characteristic probabilities |
US20130253909A1 (en) * | 2012-03-23 | 2013-09-26 | Tata Consultancy Services Limited | Second language acquisition system |
US8600730B2 (en) * | 2011-02-08 | 2013-12-03 | Microsoft Corporation | Language segmentation of multilingual texts |
US20150161996A1 (en) * | 2013-12-10 | 2015-06-11 | Google Inc. | Techniques for discriminative dependency parsing |
CN112673421A (en) * | 2018-11-28 | 2021-04-16 | 谷歌有限责任公司 | Training and/or using language selection models to automatically determine a language for voice recognition of spoken utterances |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5625749A (en) * | 1994-08-22 | 1997-04-29 | Massachusetts Institute Of Technology | Segment-based apparatus and method for speech recognition by analyzing multiple speech unit frames and modeling both temporal and spatial correlation |
US5995919A (en) * | 1997-07-24 | 1999-11-30 | Inventec Corporation | Multi-lingual recognizing method using context information |
US6085160A (en) * | 1998-07-10 | 2000-07-04 | Lernout & Hauspie Speech Products N.V. | Language independent speech recognition |
US6167377A (en) * | 1997-03-28 | 2000-12-26 | Dragon Systems, Inc. | Speech recognition language models |
US6460017B1 (en) * | 1996-09-10 | 2002-10-01 | Siemens Aktiengesellschaft | Adapting a hidden Markov sound model in a speech recognition lexicon |
-
1999
- 1999-09-02 IT IT1999TO000734A patent/IT1310154B1/en active
-
2000
- 2000-08-28 US US09/649,632 patent/US6832191B1/en not_active Expired - Lifetime
- 2000-09-01 CA CA002317231A patent/CA2317231C/en not_active Expired - Fee Related
- 2000-09-01 JP JP2000265326A patent/JP2001109490A/en active Pending
- 2000-09-01 EP EP00118987A patent/EP1081681B1/en not_active Expired - Lifetime
- 2000-09-01 DE DE60021666T patent/DE60021666T2/en not_active Expired - Lifetime
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US5625749A (en) * | 1994-08-22 | 1997-04-29 | Massachusetts Institute Of Technology | Segment-based apparatus and method for speech recognition by analyzing multiple speech unit frames and modeling both temporal and spatial correlation |
US6460017B1 (en) * | 1996-09-10 | 2002-10-01 | Siemens Aktiengesellschaft | Adapting a hidden Markov sound model in a speech recognition lexicon |
US6167377A (en) * | 1997-03-28 | 2000-12-26 | Dragon Systems, Inc. | Speech recognition language models |
US5995919A (en) * | 1997-07-24 | 1999-11-30 | Inventec Corporation | Multi-lingual recognizing method using context information |
US6085160A (en) * | 1998-07-10 | 2000-07-04 | Lernout & Hauspie Speech Products N.V. | Language independent speech recognition |
Non-Patent Citations (8)
Title |
---|
"Continuous speech recognition with neural networks and stationary-transistion-acoustic units"; Jun. 9, 1997; pp. 2107-2111. |
"In-Service adaptation of multilingual hidden-Markov-models"; Apr. 21, 1997; pp. 1451-1454. |
"Multi-Lingual Phoneme Recognition Exploiting Acoustic-Phonetic Similarities of Sounds"; Oct. 3, 1996; pp. 2195-2198. |
"Multilingual Speech Recognition For Flexible Vocabularies" Bonaventura P. et al; CSELT Consultant, Turin,IT . . . ,Dipartimento di Elettronica e Inf. Universita de Padova IT . . . ; 5 pages. |
"On-Line Adaptive Learning of the Continuous Density Hidden . . . " Qiang Huo et al; IEEE Transaction on Speech and Audio . . . ; vol. 5, No. 2, Mar. 97; 12 pages. |
"Speaker Adaptation of CDHMMs Using Bayesian Learning", Claudia Vair et al, CSELT Torino,IT; 3 pages, no date. |
A Multilingual Acoustic-Phonetic Model Based On . . . , Giorgio Micca et al, CSELT, Turin,IT, Universita di Pisa, IT . . . , no date. |
Deller et al., "Discrete-time processing of speech signals", ISBN 0-02-328301, Macillan, 1993, pp. 680-683.* * |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20020111805A1 (en) * | 2001-02-14 | 2002-08-15 | Silke Goronzy | Methods for generating pronounciation variants and for recognizing speech |
US20030040907A1 (en) * | 2001-08-24 | 2003-02-27 | Sen-Chia Chang | Speech recognition system |
US7043429B2 (en) * | 2001-08-24 | 2006-05-09 | Industrial Technology Research Institute | Speech recognition with plural confidence measures |
US20050171766A1 (en) * | 2002-02-28 | 2005-08-04 | Dario Albesano | Method for accelerating the execution of speech recognition neural networks and the related speech recognition device |
US7827031B2 (en) * | 2002-02-28 | 2010-11-02 | Loquendo S.P.A. | Method for accelerating the execution of speech recognition neural networks and the related speech recognition device |
US20050055197A1 (en) * | 2003-08-14 | 2005-03-10 | Sviatoslav Karavansky | Linguographic method of compiling word dictionaries and lexicons for the memories of electronic speech-recognition devices |
US7406408B1 (en) * | 2004-08-24 | 2008-07-29 | The United States Of America As Represented By The Director, National Security Agency | Method of recognizing phones in speech of any language |
US9202470B2 (en) * | 2009-09-28 | 2015-12-01 | Broadcom Corporation | Speech recognition using speech characteristic probabilities |
US20130151254A1 (en) * | 2009-09-28 | 2013-06-13 | Broadcom Corporation | Speech recognition using speech characteristic probabilities |
US8600730B2 (en) * | 2011-02-08 | 2013-12-03 | Microsoft Corporation | Language segmentation of multilingual texts |
US9400787B2 (en) | 2011-02-08 | 2016-07-26 | Microsoft Technology Licensing, Llc | Language segmentation of multilingual texts |
US20130253909A1 (en) * | 2012-03-23 | 2013-09-26 | Tata Consultancy Services Limited | Second language acquisition system |
US9390085B2 (en) * | 2012-03-23 | 2016-07-12 | Tata Consultancy Sevices Limited | Speech processing system and method for recognizing speech samples from a speaker with an oriyan accent when speaking english |
US20150161996A1 (en) * | 2013-12-10 | 2015-06-11 | Google Inc. | Techniques for discriminative dependency parsing |
US9507852B2 (en) * | 2013-12-10 | 2016-11-29 | Google Inc. | Techniques for discriminative dependency parsing |
CN112673421A (en) * | 2018-11-28 | 2021-04-16 | 谷歌有限责任公司 | Training and/or using language selection models to automatically determine a language for voice recognition of spoken utterances |
Also Published As
Publication number | Publication date |
---|---|
JP2001109490A (en) | 2001-04-20 |
IT1310154B1 (en) | 2002-02-11 |
EP1081681B1 (en) | 2005-08-03 |
EP1081681A3 (en) | 2001-12-19 |
CA2317231A1 (en) | 2001-03-02 |
EP1081681A2 (en) | 2001-03-07 |
DE60021666D1 (en) | 2005-09-08 |
ITTO990734A1 (en) | 2001-03-02 |
CA2317231C (en) | 2008-01-22 |
DE60021666T2 (en) | 2006-04-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US7716050B2 (en) | Multilingual speech recognition | |
CN109410914B (en) | Method for identifying Jiangxi dialect speech and dialect point | |
EP0533491B1 (en) | Wordspotting using two hidden Markov models (HMM) | |
US6243680B1 (en) | Method and apparatus for obtaining a transcription of phrases through text and spoken utterances | |
US6542866B1 (en) | Speech recognition method and apparatus utilizing multiple feature streams | |
Lin et al. | Language identification using pitch contour information | |
JP2001503154A (en) | Hidden Markov Speech Model Fitting Method in Speech Recognition System | |
US6832191B1 (en) | Process for implementing a speech recognizer, the related recognizer and process for speech recognition | |
Beyerlein et al. | Large vocabulary continuous speech recognition of Broadcast News–The Philips/RWTH approach | |
Wang et al. | A multi-pass linear fold algorithm for sentence boundary detection using prosodic cues | |
Kumpf et al. | Foreign speaker accent classification using phoneme-dependent accent discrimination models and comparisons with human perception benchmarks | |
US20230298564A1 (en) | Speech synthesis method and apparatus, device, and storage medium | |
Manjunath et al. | Development of phonetic engine for Indian languages: Bengali and Oriya | |
Parikh et al. | Gujarati speech recognition–A review | |
Manjunath et al. | Automatic phonetic transcription for read, extempore and conversation speech for an Indian language: Bengali | |
Kumaran et al. | Attention shift decoding for conversational speech recognition. | |
Pylkkönen et al. | Using phone durations in Finnish large vocabulary continuous speech recognition | |
Zgank et al. | Crosslingual transfer of source acoustic models to two different target languages | |
Pellegrino et al. | Comparison of two phonetic approaches to language identification | |
EP1589524A1 (en) | Method and device for speech synthesis | |
Mokbel et al. | Derivation of the optimal set of phonetic transcriptions for a word from its acoustic realizations | |
JP3457578B2 (en) | Speech recognition apparatus and method using speech synthesis | |
Cosi et al. | HMM/Neural Network-Based System for Italian Continuous Digit Recognition | |
Laureys et al. | An improved algorithm for the automatic segmentation of speech corpora | |
Zacharie et al. | Keyword spotting on word lattices |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: CSELT-CENTRO STUDI E LABORATORI TELECOMUNICAZIONI Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:FRASCA, ALESSANDRA;MICCA, GIORGIO;PALME, ENRICO;REEL/FRAME:011062/0598 Effective date: 20000821 |
|
AS | Assignment |
Owner name: TELECOM ITALIA LAB S.P.A., ITALY Free format text: CHANGE OF NAME;ASSIGNOR:CSELT-CENTRO STUDI E LABORATORI TELECOMUNICAZIONI S.P.A.;REEL/FRAME:012950/0117 Effective date: 20010226 |
|
STCF | Information on status: patent grant |
Free format text: PATENTED CASE |
|
FPAY | Fee payment |
Year of fee payment: 4 |
|
REMI | Maintenance fee reminder mailed | ||
AS | Assignment |
Owner name: LOQUENDO SPA, ITALY Free format text: CHANGE OF NAME;ASSIGNOR:TELECOM ITALIA LAB SPA;REEL/FRAME:023319/0355 Effective date: 20010702 |
|
FPAY | Fee payment |
Year of fee payment: 8 |
|
AS | Assignment |
Owner name: NUANCE COMMUNICATIONS, INC., MASSACHUSETTS Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:LOQUENDO S.P.A.;REEL/FRAME:031266/0917 Effective date: 20130711 |
|
FPAY | Fee payment |
Year of fee payment: 12 |